As part of the Aristotle project, UCSB constructed a system for predicting the “bid price” that an AWS user should bid for AWS spot instances—a volatile class of service offered by AWS at a reduced cost. A user who chooses the bid price get a statistical guarantee, i.e., an SLA, on the maximum price that the instance could cost moving forward into the future. The system is called DrAFTS—Durability Agreements from Time Series.
DrAFTS utilizes QBETS internally. It has been tested on individual job launches and it verifies when a user wants to run one job. Each of these experiments was a week long and launched 100 jobs (one at a time, in sequence) so, statistically it is operational for individual users with occasional needs.
At the request of the Globus Project, we collaborated with Ian Foster’s team and integrated DrAFTS with Globus Genomics. Early results are promising. DrAFTS testing and optimization continues.